A research team at Penn State developed physics-informed reinforcement learning software that can cut data center cooling costs by up to 25% without any hardware changes. The system analyzes real-time climate conditions and electricity pricing, then dynamically adjusts how aggressively a facility cools — ramping up when power is cheap and outdoor conditions are favorable, pulling back when they're not.
The case for the approach is that fixed-setpoint cooling is wasteful by design. Traditional building management systems target a constant temperature setpoint regardless of what's happening outside or what electricity costs in the next 15-minute interval. A Houston data center running its chillers at full capacity on a 45°F December morning is leaving money on the floor.
Institution: Penn State, led by Professor Wangda Zuo (architectural engineering). Method: physics-informed reinforcement learning trained in a digital twin of the facility. Test environment: simulated Houston, TX data center — high heat and humidity. Hardware component thermal limits integrated directly into the model to prevent damage. Results: up to 25% reduction in cooling electricity costs. Presentation: IEEE ITherm Conference, May 2026.
Standard reinforcement learning for building control has a known failure mode: the model learns behaviors that minimize its training objective (electricity cost) while occasionally violating physical constraints it was never explicitly taught about. A model that discovers it can reduce cooling costs by running hardware hotter than its rated thermal envelope will do exactly that until something fails.
Physics-informed reinforcement learning bakes the operating envelopes of every hardware component into the model structure itself, not as soft penalties but as hard constraints. The system cannot recommend a cooling setpoint that would push a GPU or power supply outside its rated temperature range. First author Viswanathan Ganesh noted this approach also dramatically reduces the training data required — the physics replaces data that would otherwise need to be collected empirically.
The Penn State team frames the system as "a lower-cost alternative to hardware upgrades like liquid cooling." That framing is worth examining. For existing air-cooled facilities running legacy CPU-heavy workloads, dynamic cooling optimization is a real and achievable win. For facilities running current-generation GPU hardware at 100+ kW per rack, the 25% cost reduction on cooling electricity is significant — but it does not change the fundamental physics that require liquid cooling at those densities.
The more useful framing: this technology and liquid cooling address different problems. Dynamic cooling optimization reduces the operating cost of whatever cooling system a facility already has. Liquid cooling changes which cooling system a facility can use at high rack densities. An AI-optimized liquid cooling system that reads weather data and adjusts CDU setpoints accordingly would capture both benefits. That integration is the obvious next step for anyone deploying this research at scale.
Real-time electricity pricing varies by 3x to 10x across a single day in markets like ERCOT (Texas) and PJM (Mid-Atlantic). A data center that pre-cools its thermal mass during off-peak hours — storing cold in the building structure and in chilled water tanks — and reduces cooling intensity during peak-price windows can cut electricity bills substantially without changing average temperatures. The Penn State system formalizes this intuition with a model that can operate at scale across a full facility's cooling loop.
Cooling accounts for 30 to 40% of total data center electricity consumption. A 25% reduction in cooling electricity translates to 7.5 to 10% of total facility electricity cost eliminated through software alone. For a 100 MW facility paying $0.05/kWh on average, that is $3 to $4 million annually. The capital cost of the software system is a small fraction of that payback.